An Active Learning Approach to Predict Intermolecular Non-Covalent Interactions in Organic Crystals

Intermolecular noncovalent interactions play a crucial role in the assembly of organic semiconductors (OSC). While various quantum-chemical techniques, such as symmetry-adapted perturbation theory (SAPT), are available to evaluate these interactions, they can be computationally expensive, especially for the large building blocks typical in OSC. This computational burden hinders the use of machine-driven searches across the OSC chemical and materials landscape. Machine learning (ML) models have emerged as efficient approaches to provide rapid predictions of molecular and material properties at significantly lower computational costs than quantum-chemical methods. These models, however, often rely on large, labeled datasets that can be difficult to obtain. To address this challenge, we develop an active learning ML approach that is designed to reduce the need for extensive labeled data. The active learning approach identifies areas in the chemical space where the model uncertainty is highest and enables more targeted data generation. This active learning approach is demonstrated to facilitate fast and accurate prediction of intermolecular noncovalent interactions in OSC, opening new avenues for rapid materials discovery.